That seems dubious. There are a finite number of cars you can get on the road at a given speed. If the total travel time is roughly a constant as this article implies, autonomous cars will be equivalent to adding a few more lanes, meaning there will be no long-term change.
Traffic results when car densities pass a certain threshold and the speed the cars are traveling at prevents humans from being able to react in time or appropriately.
You might have even seen this phenomenon in real life: Driving on a highway, suddenly you get stuck in traffic. As you pass through the traffic zone, there are no cars pulled over, no accident, no disruption in the quality of road. What most likely happened was that someone accidentally slammed on their brakes which caused an increase in car density. If there are enough cars that are in bound to this traffic zone before the density can disperse, the traffic zone will persist.
Also notice that traffic zones often appear in situation where there's a positive increase in elevation. What happens is that if people aren't using cruise control (or their cruise control, control system isn't tight enough) their speed will decrease as they ascend the hill, causing some congestion and as like if the density of traffic is high enough, the cars behind them will have to slow down. You may think that if the cars behind them are also ascending the hill they'll also slow down in the same way but some cars may use cruise control while others don't (resulting in congestion) or the differences in each car's torque, power-output, efficiency curves differ causing each to slow down at different rates (often leading to the least capable car governing the rate)
Use a worst-case scenario to analyze. Assume you have a network of diverless cars all linked together that maintain equidistant space between each other. If a new cars needs to join their network / pack the minimum amount of delay that the cars (behind the entering car) need to encounter is the (Length of the incoming car + gap tolerance between the other car) / (the sustained speed of the pack) =25/5280/65*3600 ~= .25 seconds per car. Now do a thought experiment that the high way you drive on was completely full (like we said in our original problem statement) do you think if you added one additional car to the highway, would everyone's (who's downstream of the merge) commute time be only increase ~.25 seconds? I'm very confident it's be much more than that.
Traffic management is a task that needs high coordination, cooperation and reaction times above what unassisted humans are capable of.
The bigger opportunity might be in traffic management. If a control system tries to get humans to do things, they will cheat at nearly every opportunity. Computer guided cars can be programmed to cooperate a little more.
If everything is networked, dynamic tolls also become easier. I guess a lot of people react badly to the concept, but it creates the ability to simply prevent heavy congestion.
